Evidence-Grounded Question Answering Task

An Evidence-Grounded Question Answering Task is a question answering task that is an explainable NLP task requiring systems to identify supporting evidence passages that justify answer span selections.

• AKA: Evidence-Based QA Task, Interpretable Question Answering.
• Context:
  • It can typically require Answer Extraction with evidence identification.
  • It can typically enforce Answer Justification through passage selection.
  • It can typically support Answer Verification via evidence inspection.
  • It can typically enable Error Attribution between wrong answers and bad evidence.
  • It can typically promote User Trust through transparent reasoning.
  • ...
  • It can often involve Multi-Passage Reasoning for complex questions.
  • It can often require Evidence Ranking by relevance scores.
  • It can often incorporate Confidence Calibration based on evidence strength.
  • It can often support Interactive QA with evidence exploration.
  • ...
  • It can range from being a Single-Evidence QA Task to being a Multi-Evidence QA Task, depending on its evidence requirement.
  • It can range from being a Extractive Evidence QA Task to being an Abstractive Evidence QA Task, depending on its answer generation method.
  • ...
  • It can process Question-Document Pairs requiring grounded answers.
  • It can output Answer Spans with supporting passages.
  • It can be solved by an Evidence-Grounded QA System.
  • It can be evaluated by an Evidence-Grounded QA Performance Measure.
  • ...
• Example(s):
  • Natural Questions Task with long answers and short answers.
  • MS MARCO Task requiring passage-level evidence.
  • SearchQA Task with evidence snippet requirement.
  • HotpotQA Task needing multi-hop evidence.
  • ELI5 Task with explanation requirements.
  • ...
• Counter-Example(s):
  • Closed-Book QA Tasks, which answer from memorized knowledge.
  • Yes/No QA Tasks, which lack evidence requirements.
  • Multiple Choice QA Tasks, which select from given options.
• See: Explainable Question Answering, Evidence-Based NLP Task, Reading Comprehension Task, Interpretable QA.

---